Angshuman Rudra, Director of Product Management at TapClicks, leading data, analytics and AI initiatives in MarTech platforms.
MIT’s State of AI In Business report highlights that 95% of AI pilots fail to deliver ROI. The report calls this the GenAI Divide—the difference between experimental adoption and systems that create measurable business value.
A major driver of this gap is the reliance on single-purpose AI agents. These agents work well in bounded domains: monitoring a campaign, drafting content or flagging anomalies. But they struggle when business processes span multiple teams, require coordination across systems or depend on continuous adaptation.
Single-Agent Systems
Enterprises often begin with single agents. A campaign monitoring agent might track media KPIs and generate alerts. A service agent may resolve simple customer requests. These agents are useful for defined, repetitive tasks.
However, they have limits. They do not scale well into environments with multiple objectives, conflicting priorities or real-time dependencies. Their decision-making scope is narrow, and the absence of coordination creates operational bottlenecks.
Multi-Agent Systems
Gartner’s analysis of AI adoption emphasizes the role of multi-agent systems (MAS). MAS is a set of specialized agents that interact to achieve broader goals. This model distributes responsibilities across agents and introduces redundancy, cross-validation and adaptability.
Examples:
Logistics: A weather agent predicts disruptions, a scheduling agent reroutes shipments, and a supplier agent manages notifications.
Finance: An anomaly detection agent identifies irregularities, a compliance agent checks regulations, and an explanation agent translates findings for audit.
Marketing: A benchmarking agent compares client results to peers, a pacing agent monitors budgets, an attribution agent maps paths, and a reporting agent assembles deliverables.
This division of labor mirrors enterprise structures—different roles working toward a shared outcome.
Why Pilots Fail
MIT’s research identifies three consistent barriers:
Lack of learning: Pilots that do not adapt to feedback lose relevance quickly.
Workflow isolation: Systems that cannot embed into existing processes are abandoned.
Overreliance on single agents: Stand-alone tools cannot handle cross-functional complexity.
Multi-agent systems address these barriers by embedding feedback loops, integrating across processes and distributing intelligence across specialized functions.
Beyond these technical challenges, many pilots fail for organizational reasons. Teams often treat AI as a short-term experiment rather than part of an operating model. Pilots are launched without clear metrics for success or without integration into production workflows and data governance processes. As a result, even when the model performs well, it doesn’t translate into business impact or sustained adoption.
Architectural Implications
According to Gartner’s How to Implement AI Agents to Transform Business Models, enterprises are expected to move toward agent-first architectures, where orchestration agents manage workflows, integrations and user interactions. While this stage is still emerging, the direction is clear. Enterprises that continue to build around disconnected single agents will re-create silos. Those that design for orchestration and coordination will have systems that scale.
Executive Actions
Select use cases carefully: Focus on cross-functional processes where MAS provide measurable value.
Assess integration: Ask whether tools learn from use, share context and coordinate with others.
Implement governance: Require audit trails, controls and human oversight where necessary.
Favor open ecosystems: Just as REST became the API standard, open protocols for agent coordination will define adoption.
Time Sensitivity
The MIT report highlights an 18-month window before early AI architecture choices become difficult to reverse. Once workflows and feedback loops are embedded, switching costs increase quickly. The critical question is not only whether to rely on isolated agents or adopt multi-agent orchestration, but also how the system is designed in either case. Early design decisions will shape scalability, governance and the ability to adapt over the long term.
Conclusion
The next phase of enterprise AI adoption will depend less on individual model performance and more on how systems are connected. Organizations that continue to deploy isolated agents will see incremental productivity gains but limited strategic impact. Sustainable value will come from architectures that enable coordination, data sharing and continuous adaptation across agents.
In the end, the GenAI Divide is less about access to AI models and more about how effectively those models are integrated into business systems.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?